Model catalogue · OpenAI
Whisper
The most widely used model for turning speech into text, covering more than 99 languages - for dictation, client conversations and meeting recordings. It produces a transcript, not an assessment of what was said; it does not detect speaker changes by itself, and during longer silences it can invent text, which is why transcripts have to be listened back against. On four routes the contractual chain is closed: on your own hardware, where no third party participates, in dedicated operation on reserved hardware, via the Open Telekom Cloud in German data centres, and via Azure, where the confidentiality undertaking under section 203 of the German Criminal Code is signed. For dictation the abuse monitoring active in the delivered state has to be considered as well.
Everything section 203 requires is settled, including the provider's confidentiality undertaking. You do not have to clarify anything further.
What it is good for
- MIT licence and small enough to run alongside on any dedicated instance
- More than 99 languages, robust even with background noise and telephone quality
- In production use for years - the most mature model in this class
Where it stops
- Does not detect speaker changes by itself - attribution in a multi-party conversation needs a second step
- Tends to invent text during silence or unintelligible passages; transcripts have to be listened back against
- No real-time streaming output - the recording is processed as a whole
Routes examined
The same model, procured differently, is a different legal position. What is examined per route is the contract chain, not the model.
| Route | Section 203 status | Version on this route |
|---|---|---|
| Dedicated instance in Germany, operated by Gosign | covered | - |
| German data centres through the Gosign chain | covered | - |
| Self-operated at the firm (open weights) | covered | - |
| Microsoft Azure | covered | - |
The conditions behind a conditional status - which piece is open on which route and who closes it - are set out in the full matrix in the German catalogue, and we go through them in the conversation. The German page for Whisper carries the sources with their retrieval dates.
Is Whisper the right fit?
We go through your tasks and the route that carries them - and say plainly where it does not.
Arrange a conversation